toolkit / e6db_reader.py
k4d3's picture
joy: use tag categories for building prompts
3821594
raw
history blame
11.9 kB
"Python only utils (no dependencies)"
import gzip
import json
import logging
import math
import warnings
from pathlib import Path
from typing import Callable, Iterable
logger = logging.getLogger(__name__)
tag_categories = [
"general",
"artist",
None, # Invalid catid
"copyright",
"character",
"species",
"invalid",
"meta",
"lore",
"pool",
]
tag_category2id = {v: k for k, v in enumerate(tag_categories) if v}
tag_categories_colors = [
"#b4c7d9",
"#f2ac08",
None, # Invalid catid
"#d0d",
"#0a0",
"#ed5d1f",
"#ff3d3d",
"#fff",
"#282",
"wheat",
]
tag_categories_alt_colors = [
"#2e76b4",
"#fbd67f",
None, # Invalid catid
"#ff5eff",
"#2bff2b",
"#f6b295",
"#ffbdbd",
"#666",
"#5fdb5f",
"#d0b27a",
]
def load_tags(data_dir):
"""
Load tag data, returns a tuple `(tag2idx, idx2tag, tag_categories)`
* `tag2idx`: dict mapping tag and aliases to numerical ids
* `idx2tag`: list mapping numerical id to tag string
* `tag_categories`: byte string mapping numerical id to categories
"""
data_dir = Path(data_dir)
with gzip.open(data_dir / "tags.txt.gz", "rt", encoding="utf-8") as fd:
idx2tag = fd.read().split("\n")
if not idx2tag[-1]:
idx2tag = idx2tag[:-1]
with gzip.open(data_dir / "tag2idx.json.gz", "rb") as fp:
tag2idx = json.load(fp)
with gzip.open(data_dir / "tags_categories.bin.gz", "rb") as fp:
tag_categories = fp.read()
logging.info(f"Loaded {len(idx2tag)} tags, {len(tag2idx)} tag2id mappings")
return tag2idx, idx2tag, tag_categories
def load_implications(data_dir):
"""
Load implication mappings. Returns a tuple `(implications, implications_rej)`
* `implications`: dict mapping numerical ids to a list of implied numerical
ids. Contains transitive implications.
* `implications_rej`: dict mapping tag strings to a list of implied
numerical ids. keys in implications_rej are tags that have a very little
usage (less than 2 posts) and don't have numerical ids associated with
them.
"""
with gzip.open(data_dir / "implications.json.gz", "rb") as fp:
implications = json.load(fp)
implications = {int(k): v for k, v in implications.items()}
with gzip.open(data_dir / "implications_rej.json.gz", "rb") as fp:
implications_rej = json.load(fp)
logger.info(
f"Loaded {len(implications)} implications + {len(implications_rej)} implication from tags without id"
)
return implications, implications_rej
def tag_rank_to_freq(rank: int) -> float:
"""Approximate the frequency of a tag given its rank"""
return math.exp(26.4284 * math.tanh(2.93505 * max(1, rank) ** (-0.136501)) - 11.492)
def tag_freq_to_rank(freq: int) -> float:
"""Approximate the rank of a tag given its frequency"""
log_freq = math.log(freq)
return math.exp(
-7.57186
* (0.0465456 * log_freq - 1.24326)
* math.log(1.13045 - 0.0720383 * log_freq)
+ 12.1903
)
InMapFun = Callable[[str, int | None], list[str]]
OutMapFun = Callable[[str], list[str]]
class TagNormalizer:
"""
Map tag strings to numerical ids, and vice versa.
Multiple strings can be mapped to a single id, while each id maps to a
single string. As a result, the encode/decode process can be used to
normalize tags to canonical spelling.
See `add_input_mappings` for adding aliases, and `rename_output` for setting
the canonical spelling of a tag.
"""
def __init__(self, path_or_data: str | Path | tuple[dict, list, bytes]):
if isinstance(path_or_data, (Path, str)):
data = load_tags(path_or_data)
else:
data = path_or_data
self.tag2idx, self.idx2tag, self.tag_categories = data
def get_category(self, tag: int | str, as_string=True) -> int:
if isinstance(tag, str):
tag = self.encode(tag)
cat = self.tag_categories[tag]
if as_string:
return tag_categories[cat]
return cat
def encode(self, tag: str, default=None):
"Convert tag string to numerical id"
return self.tag2idx.get(tag, default)
def decode(self, tag: int | str):
"Convert numerical id to tag string"
if isinstance(tag, str):
return tag
return self.idx2tag[tag]
def get_reverse_mapping(self):
"""Return a list mapping id -> [ tag strings ]"""
res = [[] for i in range(len(self.idx2tag))]
for tag, tid in self.tag2idx.items():
res[tid].append(tag)
return res
def add_input_mappings(
self, tags: str | Iterable[str], to_tid: int | str, on_conflict="raise"
):
"""Associate tag strings to an id for recognition by `encode`
`on_conflict` defines what to do when the tag string is already mapped
to a different id:
* "raise": raise an ValueError (default)
* "warn": raise a warning
* "overwrite_rarest": make the tag point to the most frequently used tid
* "overwrite": silently overwrite the mapping
* "silent", or any other string: don't set the mapping
"""
tag2idx = self.tag2idx
if not isinstance(to_tid, int):
to_tid = tag2idx[to_tid]
if isinstance(tags, str):
tags = (tags,)
for tag in tags:
conflict = tag2idx.get(tag, to_tid)
if conflict != to_tid:
msg = f"mapping {tag!r}->{self.idx2tag[to_tid]!r}({to_tid}) conflicts with previous mapping {tag!r}->{self.idx2tag[conflict]!r}({conflict})."
if on_conflict == "raise":
raise ValueError(msg)
elif on_conflict == "warn":
logger.warning(msg)
elif on_conflict == "overwrite_rarest" and to_tid > conflict:
continue
elif on_conflict != "overwrite":
continue
tag2idx[tag] = to_tid
def remove_input_mappings(self, tags: str | Iterable[str]):
"""Remove tag strings from the mapping"""
if isinstance(tags, str):
tags = (tags,)
for tag in tags:
if tag in self.tag2idx:
del self.tag2idx[tag]
else:
logger.warning(f"tag {tag!r} is not a valid tag")
def rename_output(self, orig: int | str, dest: str):
"""Change the tag string associated with an id. Used by `decode`."""
if not isinstance(orig, int):
orig = self.tag2idx[orig]
self.idx2tag[orig] = dest
def map_inputs(
self, mapfun: InMapFun, prepopulate=True, on_conflict="raise"
) -> "TagNormalizer":
tag2idx = self.tag2idx.copy() if prepopulate else {}
res = type(self)((tag2idx, self.idx2tag, self.tag_categories))
for tag, tid in self.tag2idx.items():
res.add_input_mappings(mapfun(tag, tid), tid, on_conflict=on_conflict)
return res
def map_outputs(self, mapfun: OutMapFun) -> "TagNormalizer":
idx2tag = [mapfun(t, i) for i, t in enumerate(self.idx2tag)]
return type(self)((self.tag2idx, idx2tag, self.tag_categories))
def get(self, key: int | str, default=None):
"""
Returns the string tag associated with a numerical id, or conversely,
the id associated with a tag.
"""
if isinstance(key, int):
idx2tag = self.idx2tag
if key >= len(idx2tag):
return default
return idx2tag[key]
return self.tag2idx.get(key, default)
class TagSetNormalizer:
def __init__(self, path_or_data: str | Path | tuple[TagNormalizer, dict, dict]):
if isinstance(path_or_data, (Path, str)):
data = TagNormalizer(path_or_data), *load_implications(path_or_data)
else:
data = path_or_data
self.tag_normalizer, self.implications, self.implications_rej = data
def map_inputs(self, mapfun: InMapFun, on_conflict="raise") -> "TagSetNormalizer":
tag_normalizer = self.tag_normalizer.map_inputs(mapfun, on_conflict=on_conflict)
implications_rej: dict[str, list[str]] = {}
for tag_string, implied_ids in self.implications_rej.items():
for new_tag_string in mapfun(tag_string, None):
conflict = implications_rej.get(new_tag_string, implied_ids)
if conflict != implied_ids:
msg = f"mapping {tag_string!r}->{implied_ids} conflicts with previous mapping {tag_string!r}->{conflict}."
if on_conflict == "raise":
raise ValueError(msg)
elif on_conflict == "warn":
warnings.warn(msg)
elif on_conflict != "overwrite":
continue
implications_rej[new_tag_string] = implied_ids
res = type(self)((tag_normalizer, self.implications, implications_rej))
return res
def map_outputs(self, mapfun: OutMapFun) -> "TagSetNormalizer":
tag_normalizer = self.tag_normalizer.map_outputs(mapfun)
return type(self)((tag_normalizer, self.implications, self.implications_rej))
def get_implied(self, tag: int | str) -> list[int]:
if isinstance(tag, int):
return self.implications.get(tag, ())
else:
return self.implications_rej.get(tag, ())
def encode(
self,
tags: list[str],
keep_implied: bool | set[int] = False,
max_antecedent_rank: int | None = None,
drop_antecedent_rank: int | None = None,
) -> tuple[list[int | str], set[int]]:
"""
Encode a list of string as numerical ids and strip implied tags.
Unknown tags are returned as strings.
Returns :
* a list of tag ids and unknown tag strings,
* a list of implied tag ids.
"""
tag2idx = self.tag_normalizer.tag2idx
N = len(tag2idx)
max_antecedent_rank = max_antecedent_rank or N + 1
drop_antecedent_rank = drop_antecedent_rank or N + 1
get_implied = self.implications.get
get_implied_rej = self.implications_rej.get
stack = [tag2idx.get(tag, tag) for tag in tags[::-1]]
implied = set()
res = dict() # dict as a cheap ordered set
while stack:
tag = stack.pop()
if isinstance(tag, int):
antecedent_rank = tag
consequents = get_implied(tag)
else:
# the tag might be a very rare antecedent (less than two posts)
# that doesn't have a tag id
antecedent_rank = N
consequents = get_implied_rej(tag)
if consequents:
if antecedent_rank < max_antecedent_rank:
implied.update(consequents)
else:
# The implied tags from low frequency antecedent (high rank)
# are added to the list and instead the antecedent may be
# dropped
stack.extend(consequents)
if antecedent_rank >= drop_antecedent_rank:
continue
res[tag] = None
res = res.keys()
if not keep_implied:
res = [t for t in res if t not in implied]
elif isinstance(keep_implied, set):
res = [t for t in res if t not in implied or t in keep_implied]
else:
res = list(res)
return res, implied
def decode(self, tags: Iterable[int | str]) -> list[str]:
idx2tag = self.tag_normalizer.idx2tag
return [idx2tag[t] if isinstance(t, int) else t for t in tags]